Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step4
ttt-discover-v3-qwen3-8b-ac1-s2-step4 is a machine learning model from EdwardoSunny. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
LoRA adapter (rank 32, all-linear) from the upstream-faithful local reproduction of TTT-Discover on autoresearch-bench ac1. Checkpoint saved after training step 3 (0-indexed). Strict upstream eval parity: 1100s hard k…
Downloads · 30 days
0
Access
Public
Updated Jul 9, 2026
Repo size
349 MB
Likes
0
Public
Click a slice to open those files.
.safetensors349 MB · 100%
From the Hugging Face model README
LoRA adapter (rank 32, all-linear) from the upstream-faithful local
reproduction of TTT-Discover on autoresearch-bench ac1. Checkpoint saved
after training step 3 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 3,
"progress/batch": 3,
"optim/lr": 4e-05,
"progress/done_frac": 0.08,
"puct/buffer_size": 56,
"puct/sampled_size": 8,
"puct/T": 1536,
"puct/scale_last": 0.4905733466030546,
"puct/buffer_value/mean": -1.6533351305242527,
"puct/buffer_value/std": 0.19603612123266803,
"puct/buffer_value/min": -2.000000000000007,
"puct/buffer_value/max": -1.5076871333714463,
"puct/buffer_timestep/mean": 0.7142857142857143,
"puct/buffer_timestep/std": 1.0301575072754257,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 2.0,
"puct/buffer_construction_len/mean": 2058.75,
"puct/buffer_construction_len/std": 1961.8656322308548,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.512467356471078,
"puct/sampled_value/std": 0.0023326953760160174,
"puct/sampled_value/min": -1.5138356438171245,
"puct/sampled_value/max": -1.5076871333714463,
"puct/sampled_timestep/mean": 2.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 2.0,
"puct/sampled_timestep/max": 2.0,
"puct/sampled_construction_len/mean": 1026.25,
"puct/sampled_construction_len/std": 33.8886042793149,
"puct/sampled_construction_len/min": 1000.0,
"puct/sampled_construction_len/max": 1070.0,
"time/sampling": 5210.423726320267,
"env/all/ac_tokens_per_turn": 7777.3125,
"env/all/ob_tokens_per_turn": 2855.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 3981984,
"env/all/total_ob_tokens": 1462080,
"env/all/time/sampling_mean": 257.19393996335566,
"env/all/time/sampling_max": 336.1923589706421,
"env/all/time/env_step_mean": 2402.446482351981,
"env/all/time/env_step_max": 4878.607094526291,
"env/all/reward/mean": 0.4932213197315986,
"env/all/reward/max": 0.6632881335191306,
"env/all/reward/min": 0.0,
"env/all/format": 1.0,
"env/all/format/min": 1.0,
"env/all/format/max": 1.0,
"env/all/reward": 0.4932213197315986,
"env/all/correctness": 0.80859375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.958959012280876,
"env/all/raw_score/min": 1.5076404096987763,
"env/all/raw_score/max": 100.00000000000001,
"env/all/initial_raw_score": -1.512467356471078,
"env/all/initial_raw_score/min": -1.5138356438171245,
"env/all/initial_raw_score/max": -1.5076871333714463,
"env/all/msg": "Success; raw_score=1.702332544036296",
"env/all/parsed_code": "```python\n# EVOLVE-BLOCK-START\n\"\"\"Hybrid Genetic + Adaptive Constraint Exploration with Dynamic LP Refinement for Convolution Minimization.\"\"\"\nimport time\nimport numpy as np\nfrom scipy import optimize\nimport copy\nimport random\n\nlinprog = optimize.linprog\nfrom scipy.optimize import Bounds\n\ndef get_good_direction_to_move_into(sequence):\n \"\"\"Computes a better sequence using LP with dynamic constraint selection and optimized step size.\"\"\"\n n = len(sequence)\n if n == 0:\n return None\n\n sum_seq = np.sum(sequence)\n if sum_seq <= 0.0:\n return None\n\n # Normalize sequence to have sum sqrt(2n)\n normalized_seq = [x * np.sqrt(2 * n) / sum_seq for x in sequence]\n conv = np.convolve(normalized_seq, normalized_seq)\n max_b = np.max(conv)\n\n # Select top 50 most influential positions for constraints\n top_k = min(50, len(conv))\n relevant_positions = np.argsort(conv)[::-1][:top_k]\n\n # Solve LP with dynamic constraints\n g_fun = solve_convolution_lp(normalized_seq, max_b, relevant_positions)\n if g_fun is None:\n return None\n\n sum_g = np.sum(g_fun)\n if sum_g <= 0.0:\n return None\n\n # Normalize g_fun\n normalized_g = [x * np.sqrt(2 * n) / sum_g for x in g_fun]\n\n # Adaptive line search for optimal t\n best_t = 0.0\n best_score = float('inf')\n t_low, t_high = 0.0, 1.0 # Explore full range for aggressive steps\n for _ in range(20):\n t = (t_low + t_high) / 2\n new_sequence = [(1 - t) * x + t * y for x, y in zip(sequence, normalized_g)]\n try:\n curr_score = evaluate_sequence(new_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_t = t\n if curr_score < best_score:\n # Move to higher t\n t_low = best_t\n else:\n # Move to lower t\n t_high = best_t\n except:\n break\n return [(1 - best_t) * x + best_t * y for x, y in zip(sequence, normalized_g)]\n\ndef solve_convolution_lp(f_sequence, rhs, relevant_positions=None):\n \"\"\"Solves the LP with dynamic constraint selection and adjusted bounds.\"\"\"\n n = len(f_sequence)\n if n == 0:\n return None\n\n # Default to top K positions if none are specified\n if relevant_positions is None:\n conv = np.convolve(f_sequence, f_sequence)\n top_k = min(50, len(conv))\n relevant_positions = np.argsort(conv)[::-1][:top_k]\n\n # Prepare constraints\n c = -np.ones(n) # Maximize sum_g\n a_ub = []\n b_ub = []\n\n # Add constraints for selected positions\n for k in relevant_positions:\n row = np.zeros(n)\n for i in range(n):\n j = k - i\n if 0 <= j < n:\n row[j] = f_sequence[i]\n a_ub.append(row)\n b_ub.append(rhs)\n\n # Add non-negativity constraints\n a_ub_nonneg = -np.eye(n)\n b_ub_nonneg = np.zeros(n)\n a_ub = np.vstack([a_ub, a_ub_nonneg])\n b_ub = np.hstack([b_ub, b_ub_nonneg])\n\n result = linprog(\n c,\n A_ub=a_ub,\n b_ub=b_ub,\n bounds=[(0, 1000.0) for _ in range(n)],\n options={\n \"time_limit\": 10.0,\n \"disp\": False,\n },\n )\n if result.success:\n return result.x\n return None\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n deadline = time.time() + budget_s - 20 # Reserve 20 seconds for final adjustments\n \n # Start from previous best if available\n prev = globals().get(\"GLOBAL_BEST_CONSTRUCTION\", None)\n best_sequence = []\n if np.random.rand() < 0.6 and prev is not None and len(prev) > 0:\n best_sequence = list(np.asarray(prev, dtype=float))\n else:\n # Use structured initialization with log-normal distribution\n n = 1000\n log_normal = np.random.lognormal(0.0, 0.5, size=n)\n base = np.clip(log_normal, 0.0, 1000.0)\n best_sequence = [max(0.0, base[i] * 0.5) for i in range(n)]\n \n current_sequence = best_sequence.copy()\n best_score = evaluate_sequence(best_sequence)\n\n # Adaptive Perturbation Strategy with structured exploration\n def targeted_perturb(seq):\n conv = np.convolve(seq, seq)\n max_positions = np.where(conv == np.max(conv))[0]\n if not max_positions.size:\n return perturb_sequence(seq)\n k = np.random.choice(max_positions)\n involved_elements = set()\n for i in range(len(seq)):\n j = i + k\n if j < len(seq):\n involved_elements.add(i)\n involved_elements.add(j)\n if not involved_elements:\n return perturb_sequence(seq)\n idx = np.random.choice(list(involved_elements))\n perturbation = np.random.normal(0, 0.1) * 0.3 # Larger step for exploration\n new_val = max(0.0, seq[idx] + perturbation)\n return [new_val if i == idx else seq[i] for i in range(len(seq))]\n \n def perturb_sequence(seq):\n idx = np.random.choice(len(seq))\n perturbation = np.random.normal(0, 0.05) * 0.3 # Moderate perturbation\n new_val = max(0.0, seq[idx] + perturbation)\n return [new_val if i == idx else seq[i] for i in range(len(seq))]\n \n while time.time() < deadline:\n h_function = get_good_direction_to_move_into(current_sequence)\n if h_function is None:\n # Use structured perturbation with focused exploration\n current_sequence = targeted_perturb(current_sequence)\n else:\n current_sequence = h_function\n\n try:\n curr_score = evaluate_sequence(current_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = current_sequence.copy()\n print(f\"New best: {best_score}\")\n except Exception as e:\n print(f\"Error during evaluation: {e}\")\n pass\n\n return [float(max(0.0, x)) for x in best_sequence]\n# EVOLVE-BLOCK-END\n```",
"env/all/time/policy": 257.19393996335566,
"env/all/time/policy/min": 83.08310294151306,
"env/all/time/policy/max": 336.1923589706421,
"env/all/time/env_step": 2402.446482351981,
"env/all/time/env_step/min": 0.007115602493286133,
"env/all/time/env_step/max": 4878.607094526291,
"env/all/time/reward_compute": 8.298084139823914e-07,
"env/all/time/reward_compute/min": 1.30385160446167e-07,
"env/all/time/reward_compute/max": 2.4922192096710205e-06,
"env/all/by_group/frac_mixed": 1.0,
"env/all/by_group/frac_all_good": 0.0,
"env/all/by_group/frac_all_bad": 0.0,
"advantage/mean": 0.02505502849817276,
"advantage/min": -1.0,
"advantage/max": 13.439199447631836,
"time/assemble_training_data": 5.478595495223999,
"time/kl_vs_base": 87.26551294326782,
"kl_policy_base": 0.0005950852646492422,
"time/train": 501.691002368927,
"time/save_checkpoint": 23.972317218780518,
"time/total": 5830.8483300209045
}
[2026-07-09T06:24:18+00:00] job=1812630 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T09:14:17+00:00] job=1813129 node=node-28 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T13:46:36+00:00] job=1813130 node=node-22 ngpu=6 ntrain=2 replicas=4 flash_attn=yes